Read the latest
Monday, September 28, 2026HotTea archive editionVerified 10:28 AM PDT

The lead

Four AI firms agree to testify as New York subpoenas SpaceXAI

The City Council says Anthropic, OpenAI, Google and Meta will appear on October 5. SpaceXAI received a subpoena.

Read the full story →Listen to the edition
New York City Council ↗New York City Council ↗

Listen to this edition

Prefer audio? The briefing has chapters and a full transcript.

Follow on Apple PodcastsFollow on SpotifyTranscriptAll episodes →

The briefing

The rest of the morning

4 more stories

02
War and securityHot

Nvidia puts AI agent permissions outside the agent

Nvidia detailed OpenShell 0.1.0 on September 28. The open-source software limits which files and services an agent can use. A separate supervisor checks each request against rules set by the operator.

The cited sources explain the controls, but they do not include an independent security audit. The policy checker can enforce the permissions an operator sets. That alone does not prove the rest of the system remains secure while the agent runs.

NVIDIA ↗NVIDIA on GitHub ↗
Read article →
04
Models and productsWarm

NaiveAI releases a 309-billion-parameter model with open weights

NaiveAI released Naive-N0.5-Flash with model weights and inference code under the MIT license. Its model card lists 309 billion total parameters, with 15.5 billion active for each token.

Open weights do not make this a small desktop model. The company says the compressed weights alone take roughly 315 gigabytes, before the extra memory needed to run the model. The company remains the only source for the performance figures.

NaiveAI ↗
Read article →
05
Politics and geopoliticsWarm

Washington and Beijing plan a channel for AI incidents

A September 25 White House fact sheet says the United States and China established an AI dialogue. The next exchange is due by November 2026. UPI reported the agreement on September 26.

This is a diplomatic commitment, not proof that an incident channel is operating. The statement does not set public reporting deadlines or explain which incidents must be shared.

The White House ↗UPI ↗
Read article →

Analysis

A shutdown promise means little until someone tests the stop button

New York wants AI companies to answer questions under oath. Software teams are putting access controls outside AI agents. Both efforts raise the same question: what happens when a system crosses its limits?

1

Authority

Identify the person or system allowed to stop an agent and revoke its access.

2

Enforcement

Test restrictions outside the agent, including the programs and child processes it starts.

3

Evidence

Record the attempted action, the block and the resulting system state.

The watchlist

Signals that could change the read

Company representatives answer questions under oath.
Independent teams try to defeat the documented restrictions.
Volume production begins in the first quarter of 2027.
Washington and Beijing provide working incident procedures.
Across the desks Who can stop the system?

Lawmakers seek testimony, developers add controls, and governments promise to share incident information.

New YorkTestimony promisedFour companies agreed to appear; SpaceXAI received a subpoena.
NvidiaControls outside agentsOpenShell checks an agent's access against rules set by the operator.
SingaporeTrial productionVSMC targets volume output in early 2027.
Open modelsLarge hardware billNaiveAI lists roughly 315 gigabytes for compressed model weights.

Editorial direction, not a financial index. Each signal is tied to this edition’s reporting.

Edition validated · 5 stories · 9 unique sources

About HotTea & our sources →